Research On Abnormal Behavior Detection Of Drivers And Passengers Based On Human Posture Estimation

With the improvement of the current economic level,the automobile industry is also developing rapidly,but at the same time,with it seriously threatens people’s property and safety of life in the event of frequent traffic accidents.At present,there are few applications to identify and detect the abnormal behaviors of drivers who have sudden diseases and drivers and passengers who have violent conflicts.In addition,the limited space,occlusion and lighting in the vehicle will lead to problems such as effectiveness,so that the abnormal behaviors can’t be tested accurately,and then prevent the accident from occurring.In the special scene in the car,for the three kinds of abnormal behaviors most easily to occur: driver lying on the steering wheel,fainting and reclining,and violent conflict between driver and passenger,this paper proposes a detection method for abnormal behaviors of drivers and passengers based on human posture estimation.The Alpha Pose model is used to reconstruct the behavior and posture of the people in the car,focus the key points on the attitude of the people,and filter out the interference of other factors such as background.Different recognition and detection algorithms are used for drivers and passengers respectively to detect abnormal behaviors effectively and accurately,and give early warning timely.The main work of this paper is as follows:(1)During the target extraction and pose estimation stage,the target characters in the image are detected and extracted through the YOLOv3 algorithm,and the person target area is estimated through the Alpha Pose model.This step is the basic work for the subsequent analysis and detection of abnormal behavior.(2)According to the driver’s abnormal behavior,the detection algorithm based on DBDC(Driver Behavior Detection Classification)model is proposed.Based on the gesture of the human body,it is estimated to get key points coordinates,then calculate relative position,the angle and other feature information data of behavioral posture,introduce the idea of Mobile Net V2,build a DBDC posture classification model,complete the classification of three types of behavioral posture: normal driving,lying on the steering wheel and fainting and reclining,to judge the driver’s behavior in the vehicle is abnormal or not.Through the self-built dataset in this paper,the accuracy of the algorithm is 97.59%.(3)For the abnormal behavior of the driver and passenger in the car,the spatiotemporal features are introduced based on the human posture to detect.Based on the results of the Alpha Pose attitude estimation,the human sitting posture map is constructed.The SIFT algorithm and the pyramid L-K optical flow algorithm are used to tract the optical flow changes in space feature points and time in the attitude map respectively,then calculate the displacement and kinetic energy changes of feature points,and the results are analyzed to judge the behavior in the vehicle is abnormal or not.On the data set constructed by ourselves,the recognition accuracy of the algorithm is 95.03%.The abnormal behavior detection algorithm in this article can meet the needs of effectively and accurate and warning when detecting the behavior of people in the vehicle,which has certain practical application value.